Calibration
RealWorldQA → MMMU-Pro
MMMU-Pro is estimated from RealWorldQA with a Hill curve fitted on 11 models measured on both: y = 0.0000 + (1.2000 − 0.0000)·x^6.00 / (0.77336^6.00 + x^6.00), R² = 0.61, cross-validated error 11.3 pp. It is used for 1 estimate.
| Estimated model | RealWorldQA | MMMU-Pro | Source |
|---|---|---|---|
| North Micro Vision Instruct | 62.2% | 25.6% | estimated ± 11.3 pp, low confidence |
Fitted on
| Model | RealWorldQA | MMMU-Pro |
|---|---|---|
| Claude Opus 4.5 | 77.0% | 70.6% |
| Gemini 3 Pro | 83.3% | 81.0% |
| GPT-5.2 | 83.3% | 79.5% |
| Kimi K2.5 | 81.0% | 78.5% |
| LFM2.5-VL-3B | 73.1% | 30.5% |
| LFM3-VL-2B | 79.3% | 45.7% |
| Qwen3.5 397B | 83.9% | 79.0% |
| Qwen3.6-27B | 84.1% | 75.8% |
| Qwen3.6-35B-A3B | 85.3% | 75.3% |
| Qwen3.7 Plus | 86.9% | 79.0% |
| Qwen3.8 Max | 88.0% | 82.3% |